HAL Id: hal-02427339
https://hal.inria.fr/hal-02427339
Submitted on 3 Jan 2020
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Reduced-order model for the BGK equation based on POD and optimal transport
Florian Bernard, Angelo Iollo, Sébastien Riffaud
To cite this version:
Florian Bernard, Angelo Iollo, Sébastien Riffaud. Reduced-order model for the BGK equation based on POD and optimal transport. MOREPAS 2018 - Model Reduction of Parametrized Systems IV, Apr 2018, Nantes, France. �hal-02427339�
Reduced-order model for the BGK equation based on POD and optimal transport
S. Riffaud 1,2 , F. Bernard 1,2 , A. Iollo 1,2
1
Institut de Mathématiques de Bordeaux, Université de Bordeaux, 33400 Talence, France
2
Memphis team, Inria Bordeaux Sud-Ouest, 33400 Talence, France
Contribution
We present a model describing a gas flow in both hydrodynamic and rarefied regimes. Since the computational cost of the Bhatnagar- Gross-Krook (BGK) equation can be prohibitive, a reduced-order ap- proximation is developed, leading to fast and accurate simulations.
BGK equation
The dynamics of the gas flow is described by the BGK equation:
∂f
∂t (x, ξ, t) + ξ · ∇
xf (x, ξ, t) = M
f(x, ξ, t) − f (x, ξ, t)
τ (x, t) (1)
where f is the density distribution function representing the density of gas particles at point x ∈ R
3, velocity ξ ∈ R
3and time t ∈ R . The Maxwellian distribution function M
fis in dimensionless form
M
f(x, ξ, t) = ρ(x, t)
(2πT (x, t))
32exp
− kξ − U (x, t)k
22T (x, t)
where ρ ∈ R is the density, U ∈ R
3is the macroscopic velocity, T ∈ R is the temperature and E ∈ R is the total energy of the gas.
Reduced-order approximation
The distribution functions are approximated by f e (x, ξ, t) =
Npod
X
n=1
a
fn(x, t)Φ
n(ξ )
and
M f
f(x, ξ, t) =
Npod
X
n=1
a
Mn(x, t)Φ
n(ξ )
where the basis functions Φ
nare built offline by Proper Orthog- onal Decomposition (POD) and the coefficients a
nare computed online by the Galerkin method.
References
[1] P. L. Bhatnagar, E. P. Gross and M. Krook. A model for collision processes in gases. I. Small amplitude processes in charged and neutral one-component systems. Physical Review, 1954.
[2] F. Bernard. Efficient asymptotic preserving schemes for BGK and ES-BGK models on cartesian grids. PhD thesis, 2015.
[3] F. Bernard, A. Iollo and S. Riffaud. Reduced-order model for the BGK equation based on POD and optimal transport. Journal of Computational Physics, 2018.
Offline phase
In the offline phase, the BGK equation (1) is sampled to collect infor- mation on the distribution functions that we want to approximate.
High-fidelity simulations provide snapshots of both the density dis- tribution function and of the Maxwellian distribution function:
S
hf= n
f (x
i, ξ, t
k) o
1≤i≤Nx 1≤k≤Nt
[ n M
f(x
i, ξ, t
k) o
1≤i≤Nx 1≤k≤Nt
Then, optimal transport provides addi- tional low-fidelity snapshots by interpolating the snapshots of S
hfto complete the sampling:
S = S
hf∪ S
lfFinally, the basis functions Φ
nare built by POD to have the best approximation in the least squares sense of the snapshots s
l∈ S :
minimize
Φ1(ξ)...ΦNpod (ξ)
Nsnaps
X
l=1
Z
R3
s
l(ξ ) − P [s
l](ξ )
2dξ
subject to
Z
R3
Φ
n(ξ )Φ
m(ξ ) dξ = δ
n,mwhere P [s
l] is the projection of s
lonto the subspace spanned by the basis functions Φ
ni.e. P [s
l](ξ ) =
Npod
P
n=1
R
R3
s
l(ξ
0)Φ
n(ξ
0) dξ
0Φ
n(ξ ) .
Online phase
During the online phase, the offline knowledge is used to compute approximations of the distribution functions at low cost. In the Galerkin method, the BGK equation (1) is projected onto the basis functions Φ
n, leading to an hyperbolic system of partial differen- tial equations:
∂a
f∂t (x, t)+A ∂a
f∂x (x, t)+ ˚ A ∂a
f∂y (x, t)+
?
A ∂a
f∂z (x, t) = a
M(x, t) − a
f(x, t) τ (x, t)
where a = (a
1, a
2, . . . , a
Npod)
T, A
n,m= R
R3
ξ
uΦ
n(ξ )Φ
m(ξ ) dξ , A ˚
n,m= R
R3
ξ
vΦ
n(ξ )Φ
m(ξ ) dξ and A
?n,m= R
R3
ξ
wΦ
n(ξ )Φ
m(ξ ) dξ . These equations are decoupled by linear changes of variables and are solved by an IMEX Runge-Kutta scheme in time and a finite volume scheme in space. To improve the accuracy of the model, the coefficients a
Mnare computed by constrained projection:
minimize
aM1 (x,t)...aM
Npod (x,t)
Z
R3
M
f(x, ξ, t) − M f
f(x, ξ, t)
2dξ
subject to
Z
R3
M f
f(x, ξ, t)
1 ξ
kξk2 2